Google Just Bought Its Way Deeper Into Custom AI Chips

Google has agreed a deal that could give it a stake in Marvell reported at 12.2 billion dollars. Marvell is not a household name. It is one of the few firms that can take a chip design from concept to a working, manufacturable accelerator with the interconnect to match.

Why Marvell, why now

Designing a good AI chip is difficult. Building the surrounding system is harder: the serialiser and deserialiser blocks, the optical links, the memory controllers, the packaging. Marvell does that work for hyperscalers. Locking in a stake means locking in engineering capacity that everyone else is also chasing.

Google has run TPUs internally for a decade. For most of that time they were a private edge, quietly cutting the cost of search ranking, translation and later Gemini. The shift over the past year has been commercial. TPUs are increasingly sold as capacity anyone can rent, and that only works if supply is dependable.

The Nvidia question

Nobody is displacing Nvidia this year. CUDA is a moat built over fifteen years and nearly every research codebase assumes it. But not every workload needs the most flexible chip on the market. Steady, well understood inference at enormous volume is exactly where custom silicon earns its keep, because you can trade generality for cost per token.

If a serious share of inference migrates to TPUs, Trainium and similar parts, Nvidia keeps the frontier training market and loses some of the boring, high volume revenue underneath it. That is the pressure point.

What it means for everyone else

For developers, more silicon competition eventually means cheaper inference, though the savings arrive slowly and usually as better free tiers rather than lower list prices. For anyone building on a single provider's stack, it is another reminder to keep your model calls behind an abstraction you control.

For the market, this is a signal about where the value is moving. The interesting companies are no longer just the ones making models. They are the ones making the machines and the links between them.

What to watch

Whether Google publishes third party TPU pricing that undercuts GPU instances on like for like inference. Whether Marvell's other hyperscaler customers get nervous about a competitor holding a stake. And whether the regulators in Brussels and Washington take an interest in a cloud provider buying into its own supplier.